Companies produced substantially more code after adopting AI coding agents, but a new study found little evidence that they completed more software features. The apparent productivity gain was absorbed by review and revision work downstream.
Harvard researchers Fiona Chen and James Stratton analyzed 300 million work events across more than 700 software firms and 700,000 employees from 2021 through March 2026. They used measured AI activity and GitHub signals to estimate when organizations introduced coding assistants or autonomous agents, then compared changes over time.
Agent adoption was associated with 30 percent more lines of code, 20 percent more commits and 23 percent more pull requests. Completion rates for Jira issues and larger epics did not change significantly. Average time from pull-request submission to merge rose 49 percent, the share requiring changes nearly doubled and review comments increased 35 percent.
More employees also participated in review, rising by 14 percent, while the researchers found no significant employment change attributable to the tools. The observational design cannot prove that agents alone caused every difference, and results may vary by team and workflow. It does indicate that code generation is not the same as delivered software: review capacity, requirements and integration remain constraints even when typing becomes faster.